Domain Adaptation Methods for Lab-to-Field Human Context Recognition.

Area of Science:

• Computer Science
• Machine Learning
• Human-Computer Interaction

Background:

• Human Context Recognition (HCR) is vital for Context-Aware (CA) applications in healthcare and security.
• Supervised HCR models trained on scripted smartphone data perform poorly on realistic, in-the-wild datasets due to data issues.
• Existing lab-to-field approaches aim to bridge this performance gap by adapting models to noisy, real-world data.

Purpose of the Study:

• To introduce Triple-DARE, a novel lab-to-field neural network method for enhancing HCR performance on in-the-wild datasets.
• To improve the robustness and accuracy of HCR models in realistic, unscripted environments.
• To address challenges like data imbalance and label noise in real-world HCR data.

Main Methods:

• Developed Triple-DARE, a domain adaptation technique utilizing three loss functions: domain alignment, classification, and joint fusion triplet loss.
• Focused on enhancing intra-class compactness and inter-class separation within the embedding space for multi-labeled datasets.
• Employed a lab-to-field strategy to learn robust representations from high-fidelity scripted data for application to noisy in-the-wild data.

Main Results:

• Triple-DARE achieved a 6.3% higher F1-score and 4.5% higher classification accuracy compared to state-of-the-art HCR baselines.
• Demonstrated significant improvements over non-adaptive HCR models, with 44.6% and 10.7% higher performance.
• Successfully enhanced HCR model performance by learning domain-invariant embeddings and preserving task-discriminative features.

Conclusions:

• Triple-DARE offers a significant advancement in domain adaptation for HCR, particularly for challenging in-the-wild datasets.
• The proposed method effectively bridges the performance gap between scripted and realistic HCR data.
• This research provides a robust solution for improving the reliability of CA applications in real-world settings.